Re: "Dying CPU not properly vacated" splat

["Paul E. McKenney" <paulmck@kernel.org>]

On Tue, Apr 26, 2022 at 03:48:06PM +0100, Valentin Schneider wrote:
> On 25/04/22 17:03, Paul E. McKenney wrote:
> > On Mon, Apr 25, 2022 at 10:59:44PM +0100, Valentin Schneider wrote:
> >> On 25/04/22 10:33, Paul E. McKenney wrote:
> >> >
> >> > So what did rcu_torture_reader() do wrong here?  ;-)
> >> >
> >>
> >> So on teardown, CPUHP_AP_SCHED_WAIT_EMPTY->sched_cpu_wait_empty() waits for
> >> the rq to be empty. Tasks must *not* be enqueued onto that CPU after that
> >> step has been run - if there are per-CPU tasks bound to that CPU, they must
> >> be unbound in their respective hotplug callback.
> >>
> >> For instance for workqueue.c, we have workqueue_offline_cpu() as a hotplug
> >> callback which invokes unbind_workers(cpu), the interesting bit being:
> >>
> >>                 for_each_pool_worker(worker, pool) {
> >>                         kthread_set_per_cpu(worker->task, -1);
> >>                         WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
> >>                 }
> >>
> >> The rcu_torture_reader() kthreads aren't bound to any particular CPU are
> >> they? I can't find any code that would indicate they are - and in that case
> >> it means we have a problem with is_cpu_allowed() or related.
> >
> > I did not intend that the rcu_torture_reader() kthreads be bound, and
> > I am not seeing anything that binds them.
> >
> > Thoughts?  (Other than that validating any alleged fix will be quite
> > "interesting".)
> 
> IIUC the bogus scenario is is_cpu_allowed() lets one of those kthreads be
> enqueued on the outgoing CPU *after* CPUHP_AP_SCHED_WAIT_EMPTY.teardown() has
> been run, and hilarity ensues.
> 
> The cpu_dying() condition should prevent a regular kthread from getting
> enqueued there, most of the details have been evinced from my brain but I
> recall we got the ordering conditions right...
> 
> The only other "obvious" thing here is migrate_disable() which lets the
> enqueue happen, but then balance_push()->select_fallback_rq() should punt
> it away on context switch.
> 
> I need to rediscover those paths, I don't see any obvious clue right now.

Thank you for looking into this!

The only thought that came to me was to record that is_cpu_allowed()
returned true do to migration being disabled, and then use that in later
traces, printk()s or whatever.

My own favorite root-cause hypothesis was invalidated by the fact that
is_cpu_allowed() returns cpu_online(cpu) rather than just true.  ;-)

							Thanx, Paul